Abstract

Info

Title:: Application of improved YOLOv4 algorithm in the detection of pulmonary tuberculosis

Author(s):: WANG Jinghua; YUAN Jinli; GUO Zhitao; WANG Jiahao; School of Electronic and Information Engineering, Hebei University of Technology, Tianjin 300401, China

Keywords:: Keywords: pulmonary tuberculosis deep learning feature fusion coordinate attention YOLOv4

Abstract:: Abstract: Aiming at the problem of low detection accuracy of pulmonary tuberculosis caused by the complex and large scale changes of tuberculosis lesions in CT images, YOLOv4 with an improved feature fusion block is proposed for the detection of pulmonary tuberculosis. Scale-equalizing pyramid convolution is used to capture the interaction between feature layers of different scales, and on this basis, the conflict information at different scales is filtered out by scale-equalizing adaptive spatial pyramid convolution, so as to achieve feature fusion effectively. In addition, coordinate attention is introduced on the low-level features for further improving the detection accuracy of small targets. A standardized tuberculosis CT data set is built using the information of 300 cases provided by Beijing Chest Hospital, and the experiments are conducted on the constructed data set. The input image resolution is set to 512×512. The results show that the proposed network increases mAP by 4.96% as compared with the original YOLOv4, and that it is better than the existing mainstream tuberculosis detection algorithms, such as Faster R_CNN, SSD, RetinaNet, etc. The improved YOLOv4 algorithm can effectively solve the problems of detection target scale changes and small target detection, thereby improving detection accuracy.